Google DeepMind shipped Gemini Robotics ER 2 today, July 30, 2026, and buried inside the announcement is a stat that tells you more than the demo videos do: moment-finding accuracy jumped to 91.3%, with a mean absolute distance of 0.96 seconds, at four times the execution speed of the previous generation. That's the difference between a robot that watches a task and a robot that knows, mid-motion, exactly when to stop pouring the coffee.

Gemini Robotics ER 2 is Google's new embodied reasoning model, the "high-level brain" that sits above a robot's low-level motor controllers. It doesn't move the arm itself. It watches video, decides what happens next, calls tools, and hands execution off to a vision-language-action model or a human operator. It replaces Gemini Robotics-ER 1.6, which shipped in April 2026, and it launched alongside two sibling models: Gemini Robotics 2, a full-body VLA controller, and Gemini Robotics On-Device 2, an offline version built for robots with no internet connection.

If you build with robots, evaluate physical AI vendors, or just track where Google is pointing its compute, here's what actually changed, what the benchmark numbers mean in practice, and where this fits against everything DeepMind has shipped in the Gemini Robotics line since March 2025.

Think of it less as a robot controller and more as a project manager standing next to the robot. Gemini Robotics ER 2 takes in continuous video, audio, and text, reasons about the physical scene, plans a multi-step task, and issues commands to whatever low-level system does the actual moving; a VLA model, a movement-control API, or a teleoperated human. Because it runs on the Gemini Live API's bidirectional streaming endpoint, it can "think" about the next step while the current one is still executing, instead of the stop-and-think pauses that plagued earlier robotics agents.

Three capabilities separate this release from ER 1.6:

Continuous progress classification. The model watches a video feed and buckets task progress into five bands (0-20%, 20-40%, and so on), frame by frame. That sounds mundane until you realize it's what lets a robot retry a single failed step instead of restarting an entire workflow. Google reports 57.4% accuracy here, which reads low until you compare it against the field: the company says it beats both its own prior generation and competing frontier models on the same test.

Precision moment-finding. This is the coffee-pouring example. Moment-finding means identifying the exact frame where something critical happens, so the robot knows when to stop, switch tasks, or flag a problem. At 91.3% accuracy and a 0.96-second mean absolute distance, Gemini Robotics ER 2 is landing close to models several times its size, according to Google, while running four times faster.

Multi-robot collaboration. This is new to the ER line. Gemini Robotics ER 2 can coordinate two different robot bodies on a shared task through common semantic understanding rather than hardcoded handoff logic. Google's own demo pairs Apptronik's Apollo 2 humanoid with a Franka F3 Duo arm.

Version numbers lie. "ER 2" sounds like an incremental bump from "ER 1.6," but the underlying shift is structural. Every prior Gemini Robotics-ER release, including 1.5 and 1.6 back in April 2026, reasoned over static images or short clips. Gemini Robotics ER 2 reasons over continuous video streams and tracks its own task state across minutes, not seconds. That's the actual delta, and it's the reason DeepMind paired it with a name change instead of calling it "ER 1.7."

The honest caveat: a 57.4% score on progress classification is not a number you'd want running unsupervised on a factory line yet. Google frames it as outperforming the field, and based on what's publicly benchmarked, that claim holds; but "best available" and "reliable enough to trust alone" are different bars, and the gap between them is exactly what the next 12 months of robotics coverage will be about.

No single source, including Google's own two blog posts on this launch, lays out the entire model family in one place. Here's every release from March 2025 through today, merged from the developer blog, the DeepMind blog, and the Gemini API changelog.

Two things jump out once it's laid side by side. First, Gemini Robotics ER 2 is the most openly accessible model in the entire family right now; it's the only one on general developer access via the Gemini API rather than gated behind a partner waitlist. Second, the "ER" line and the VLA line have been on separate release cadences since March 2025 and only just synced back up on the same launch day, which is itself a signal DeepMind is treating July 30 as a coordinated platform release rather than a single model drop.

Sixteen months, seven named model releases, and a consistent pattern: DeepMind ships the embodied-reasoning brain slightly ahead of or alongside the motor-control body, then widens access to the reasoning model first. If that pattern holds, expect Gemini Robotics 2 (the VLA) and On-Device 2 to move from partner waitlist to general access before the end of the year, mirroring how ER 1.6 preceded today's wider VLA release by roughly three and a half months.

Google says Gemini Robotics ER 2 is its safest robotics model yet, built on gains in Safety Instruction Following and Human Proximity benchmarks. In practice, that means the model can halt a humanoid robot when it detects a person nearby and resume only once the area clears, a behavior demonstrated in the DeepMind blog post. The company also introduced a new benchmark called ASIMOV-Agentic, which tests whether the model refuses unsafe tool calls from a VLA and knows when to ask a human for clarification instead of guessing.

This matters more than it sounds. A model that's merely accurate at spatial reasoning can still be dangerous if it doesn't know the limits of its own certainty. ASIMOV-Agentic is explicitly testing for that gap, not just raw task accuracy, which is a more mature safety framing than most robotics benchmarks used a year ago.

Developers can access the model today through the Gemini API and Google AI Studio, with private preview access on the Gemini Enterprise Agent Platform. Google has published getting-started examples on GitHub, including a live demo pairing the model with Boston Dynamics' Spot robot to fetch objects on voice command using Spot's existing movement and manipulator APIs. To build an agentic setup, you declare your low-level control interfaces, whether that's a VLA model or a movement-control API, as callable tools, then stream multimodal input directly into the model over the Live API's bidirectional endpoint.

Is Gemini Robotics ER 2 available to the public? Yes. It's live on the Gemini API and Google AI Studio as of July 30, 2026, with private preview access on the Gemini Enterprise Agent Platform.

Does Gemini Robotics ER 2 control the robot's motors directly? No. It's the reasoning layer. Motor execution is handled by a separate vision-language-action model like Gemini Robotics 2, or by a human teleoperator.

What replaced Gemini Robotics-ER 1.6? Gemini Robotics ER 2, released July 30, 2026, roughly three and a half months after ER 1.6 shipped in April.

Can two different robots use Gemini Robotics ER 2 together? Yes, this is the multi-robot collaboration feature introduced in this release, demonstrated pairing an Apptronik Apollo 2 humanoid with a Franka F3 Duo arm.

What's the moment-finding accuracy number in plain terms? 91.3%, with the model landing within roughly one second of the exact correct frame on average, at four times the speed of the prior generation.

Google says it's pushing these models toward longer, more complex task sequences next. Given the pace of the last 16 months, roughly one release every three to four months, the more useful question isn't whether the next update lands, it's whether the accuracy numbers climb fast enough to close the gap between "best in category" and "reliable enough to trust alone in a room with people."